Joint Optimization of Confidence Thresholds and Resource Allocation for Cooperative Inference in Mobile Edge Computing Systems

Abstract

To address the limitations of standalone inference, which relies exclusively on either an edge device or a server, we propose a novel cooperative inference method for mobile edge computing (MEC) systems. Our approach uses dual confidence thresholds on the edge device, equipped with a small neural network (NN) model, to filter ambiguous input images. These images are subsequently transmitted to the edge server with a large NN model and reevaluated for a final decision. We analyze the proposed method in terms of inference accuracy, delay, and energy consumption, considering the distribution of confidence scores from both positive and negative images that may trigger false alarms. Subsequently, we formulate a joint optimization problem to determine the optimal confidence thresholds at both the device and server, as well as the device's transmit power and duty cycle for transmission and reception, aiming to minimize the weighted sum of delay and energy consumption while maintaining the required accuracy level. To solve this problem, we propose a low-complexity algorithm that integrates an optimization method for finding the transmit power and duty cycle, along with a greedy search algorithm to determine effective confidence thresholds. Experimental results reveal a trade-off between inference accuracy and energy-delay cost depending on the confidence thresholds. Moreover, the joint optimization of confidence thresholds and radio resources significantly reduces delay and energy consumption while satisfying the required accuracy. Consequently, the proposed cooperative inference achieves higher accuracy than device-only inference and incurs much lower costs than server-only inference in various MEC environments. © 1967-2012 IEEE.